code/scripts/vis_sampling_v3.py

"""
Compare sampling: No-CFG (direct mask) vs CFG vs Validation pipeline
"""
import sys
sys.path.insert(0, "/data/sichengli/Code/PixelGen")

import torch
import numpy as np
from PIL import Image, ImageDraw, ImageFont
import torchvision.transforms as transforms
import torchvision.transforms.functional as TF
import os, random

from src.models.transformer.JiT_medical import JiTMedical
from src.diffusion.flow_matching.scheduling import LinearScheduler


device = torch.device("cuda:0")

# Load model
print("Loading model...")
ckpt_path = "/data/sichengli/Code/PixelGen/medical_workdirs/exp_PixelGen_Medical_B16/epoch=236-step=100000.ckpt"
ckpt = torch.load(ckpt_path, map_location="cpu", weights_only=False)
state_dict = ckpt["state_dict"]

model = JiTMedical(
    input_size=256, patch_size=16, in_channels=3,
    hidden_size=768, depth=12, num_heads=12, mlp_ratio=4.0,
    attn_drop=0.0, proj_drop=0.1, num_classes=1,
    use_bottleneck=True, bottleneck_dim=128,
    in_context_len=32, in_context_start=4, mask_in_channels=1
)
ema_state = {}
for k, v in state_dict.items():
    if k.startswith("ema_denoiser."):
        new_k = k.replace("ema_denoiser.", "").replace("_orig_mod.", "")
        ema_state[new_k] = v
model.load_state_dict(ema_state, strict=False)
model = model.to(device).eval().to(torch.float32)
print(f"Loaded EMA model ({len(ema_state)} keys)")

scheduler = LinearScheduler()


def shift_respace_fn(t, shift=1.0):
    return t / (t + (1 - t) * shift)


@torch.no_grad()
def sample_no_cfg(model, noise, mask, num_steps=50, t_eps=0.05):
    """Single-path sampling: directly pass mask to model, no CFG."""
    batch_size = noise.shape[0]
    timesteps = torch.linspace(0.0, 1 - 1.0 / num_steps, num_steps)
    timesteps = torch.cat([timesteps, torch.tensor([1.0])], dim=0)
    timesteps = shift_respace_fn(timesteps, 1.0).to(noise.device)

    y = torch.zeros(batch_size, dtype=torch.long, device=noise.device)

    x = noise
    for i in range(len(timesteps) - 1):
        t_cur = timesteps[i]
        t_next = timesteps[i + 1]
        dt = t_next - t_cur
        t_batch = t_cur.repeat(batch_size)

        # Single forward pass with mask
        pred_img = model(x, t_batch, y, mask=mask)

        # Convert x-prediction to velocity
        v = (pred_img - x) / (1.0 - t_batch.view(-1, 1, 1, 1)).clamp_min(t_eps)

        # Euler step
        x = x + v * dt

    return x


@torch.no_grad()
def sample_with_cfg(model, noise, mask, num_steps=50, cfg_scale=2.0, t_eps=0.05):
    """Dual-path CFG sampling."""
    batch_size = noise.shape[0]
    timesteps = torch.linspace(0.0, 1 - 1.0 / num_steps, num_steps)
    timesteps = torch.cat([timesteps, torch.tensor([1.0])], dim=0)
    timesteps = shift_respace_fn(timesteps, 1.0).to(noise.device)

    y = torch.zeros(batch_size, dtype=torch.long, device=noise.device)

    x = noise
    for i in range(len(timesteps) - 1):
        t_cur = timesteps[i]
        t_next = timesteps[i + 1]
        dt = t_next - t_cur
        t_batch = t_cur.repeat(batch_size)

        # CFG: concat uncond + cond
        cfg_x = torch.cat([x, x], dim=0)
        cfg_t = t_batch.repeat(2)
        cfg_y = torch.cat([y, y], dim=0)
        cfg_mask = torch.cat([torch.zeros_like(mask), mask], dim=0)

        pred = model(cfg_x, cfg_t, cfg_y, mask=cfg_mask)
        pred_v = (pred - cfg_x) / (1.0 - cfg_t.view(-1, 1, 1, 1)).clamp_min(t_eps)

        # CFG
        v_uncond, v_cond = pred_v.chunk(2)
        v = v_uncond + cfg_scale * (v_cond - v_uncond)

        x = x + v * dt

    return x


@torch.no_grad()
def sample_no_mask(model, noise, num_steps=50, t_eps=0.05):
    """No-mask sampling (like validation pipeline)."""
    batch_size = noise.shape[0]
    timesteps = torch.linspace(0.0, 1 - 1.0 / num_steps, num_steps)
    timesteps = torch.cat([timesteps, torch.tensor([1.0])], dim=0)
    timesteps = shift_respace_fn(timesteps, 1.0).to(noise.device)

    y = torch.zeros(batch_size, dtype=torch.long, device=noise.device)

    x = noise
    for i in range(len(timesteps) - 1):
        t_cur = timesteps[i]
        t_next = timesteps[i + 1]
        dt = t_next - t_cur
        t_batch = t_cur.repeat(batch_size)

        # No mask -> mask_emb = zeros
        pred_img = model(x, t_batch, y, mask=None)
        v = (pred_img - x) / (1.0 - t_batch.view(-1, 1, 1, 1)).clamp_min(t_eps)
        x = x + v * dt

    return x


# Load 6 samples
data_root = "/data2/sichengli/Data/test/Segmentation/OCTA500"
img_dir = os.path.join(data_root, "images")
mask_dir = os.path.join(data_root, "masks")
all_files = sorted([f for f in os.listdir(img_dir) if f.endswith(".png") and not f.startswith("thumb")])
random.seed(456)
selected = random.sample(all_files, 6)

images_list, masks_list = [], []
for fname in selected:
    img = Image.open(os.path.join(img_dir, fname)).convert("L")
    img = TF.resize(img, (256, 256))
    images_list.append(TF.to_tensor(img).repeat(3, 1, 1))
    mask = Image.open(os.path.join(mask_dir, fname)).convert("L")
    mask = TF.resize(mask, (256, 256), interpolation=transforms.InterpolationMode.NEAREST)
    masks_list.append(TF.to_tensor(mask))

real_images = torch.stack(images_list)
masks_tensor = torch.stack(masks_list).to(device)

# Same noise for all
torch.manual_seed(42)
shared_noise = torch.randn(6, 3, 256, 256, device=device)

# Generate with different methods
print("1/3 No-CFG (direct mask)...")
gen_no_cfg = sample_no_cfg(model, shared_noise.clone(), masks_tensor).clamp(-1, 1) * 0.5 + 0.5

print("2/3 CFG=2.0 (current)...")
gen_cfg = sample_with_cfg(model, shared_noise.clone(), masks_tensor, cfg_scale=2.0).clamp(-1, 1) * 0.5 + 0.5

print("3/3 No mask (like val pipeline)...")
gen_no_mask = sample_no_mask(model, shared_noise.clone()).clamp(-1, 1) * 0.5 + 0.5

results = {
    "Mask": None,
    "No-CFG\n(+mask)": gen_no_cfg.cpu(),
    "CFG=2.0\n(+mask)": gen_cfg.cpu(),
    "No-Mask\n(val mode)": gen_no_mask.cpu(),
    "Real": None,
}

# Create comparison grid
col_labels = list(results.keys())
n_rows = 6
n_cols = len(col_labels)
h, w = 256, 256
pad = 4
label_h = 48

canvas_w = n_cols * w + (n_cols + 1) * pad
canvas_h = n_rows * h + (n_rows + 1) * pad + label_h
canvas = np.ones((canvas_h, canvas_w, 3), dtype=np.uint8) * 30

# Column labels
try:
    font = ImageFont.truetype("/usr/share/fonts/truetype/dejavu/DejaVuSans-Bold.ttf", 16)
except Exception:
    font = ImageFont.load_default()

pil_canvas = Image.fromarray(canvas)
draw = ImageDraw.Draw(pil_canvas)
for col, label in enumerate(col_labels):
    x_pos = pad + col * (w + pad) + w // 2
    lines = label.split("\n")
    for li, line in enumerate(lines):
        bbox = draw.textbbox((0, 0), line, font=font)
        text_w = bbox[2] - bbox[0]
        draw.text((x_pos - text_w // 2, 4 + li * 20), line, fill=(255, 255, 255), font=font)
canvas = np.array(pil_canvas)

color_map = {0: (0,0,0), 50: (255,80,80), 100: (80,255,80), 150: (80,80,255), 200: (255,255,80), 250: (255,80,255)}

for row in range(n_rows):
    y = label_h + pad + row * (h + pad)

    for col_idx, col_name in enumerate(col_labels):
        x = pad + col_idx * (w + pad)

        if "Mask" == col_name:
            m = masks_tensor[row, 0].cpu().numpy()
            m_uint8 = (m * 255).astype(np.uint8)
            m_colored = np.zeros((h, w, 3), dtype=np.uint8)
            for val, color in color_map.items():
                mask_region = np.abs(m_uint8.astype(int) - val) < 13
                m_colored[mask_region] = color
            canvas[y:y+h, x:x+w] = m_colored
        elif "Real" in col_name:
            r = real_images[row].permute(1, 2, 0).numpy()
            canvas[y:y+h, x:x+w] = (r * 255).clip(0, 255).astype(np.uint8)
        else:
            g = results[col_name][row].permute(1, 2, 0).numpy()
            canvas[y:y+h, x:x+w] = (g * 255).clip(0, 255).astype(np.uint8)

out = "/data/sichengli/Code/PixelGen/medical_workdirs/exp_PixelGen_Medical_B16/val_samples/sampling_method_compare.png"
Image.fromarray(canvas).save(out)
print(f"\nSaved: {out}")